FALSE ALARM FILTERING IN A VISION TRAFFIC SIGN

RECOGNITION SYSTEM

An Approach based on AdaBoost and Heterogeneity of Texture

Sergio Lafuente-Arroyo, Saturnino Maldonado-Basc

Hilario G

omez-Moreno and Pedro Gil-Jim

enez

Deparment of Signal Theory and Communications, University of Alcal

a, Alcal

a de Henares, Madrid, Spain

Keywords:

Intelligent transportation system (ITS), Trafﬁc sign setection and recognition system (TSDRS), AdaBoost,

Support vector machines (SVMs), Pattern recognition.

Abstract:

The high variability of road sign appearance and the variety of different classes have made the recognition

of pictograms a high computational load problem in trafﬁc sign detection based on computer vision. In this

paper false alarms are reduced signiﬁcantly by designing a cascade ﬁlter based on boosting detectors and a

generative classiﬁer based on heterogeneity of texture. The false alarm ﬁlter allows us to discard many false

positives using a reduced selection of features, which are chosen from a wide set of features. Filtering is

deﬁned as a binary problem, where all speed limit signs are grouped together against noisy examples and it is

the previous stage to the input of a recognition module based on Support Vector Machines (SVMs). In a trafﬁc

sign recognition system, the number of candidate blobs detected is, in general, much higher than the number

of trafﬁc signs. As asymmetry is an inherent problem, we apply a different treatment for false negatives (FN)

and false positives (FP). The global ﬁlter offers high accuracy. It achieves very low false alarm ratio with low

computational complexity.

1 INTRODUCTION

Trafﬁc sign detection and recognition tasks based on

vision systems and artiﬁcial techniques arise in a

wide range of applications, such as intelligent vehi-

cles that operate automatically, advanced driver as-

sistance systems (ADA), which use the information

captured from sensors surrounding the vehicle to as-

sist the driver or automatic trafﬁc sign inventory in

order to maintain highway infrastructure periodically,

which allows to alert about the lack of visibility of

speciﬁc trafﬁc signs.

Independently of the application, time optimiza-

tion is a crucial item in a trafﬁc sign detection sys-

tem and one of the major problems is the high number

of false alarms that the system considers as potential

signs even when the number of positive targets in each

image is too low. As the computational load depends

on the number of samples to analyze, processing time

becomes considerable. In this paper we present a false

alarm ﬁlter integrated in a vision system in order to

discard these candidate objects as soon as possible.

There are several approaches to detect trafﬁc

signs. Many of them are based on color segmentation

using different spaces, such as RGB (de la Escalera

et al., 1997) or HSI (Maldonado-Basc

on et al., 2007)

considering particular colors of the signs present in

cluttered scenes. Other alternatives use borders detec-

tion, texture detection and genetic algorithms (de la

Escalera et al., 2003). More recently (Bahlmann

et al., 2005) suggested AdaBoost for detection. With

respect to identiﬁcation module, neuronal networks,

(Fang et al., 2003), and support vector machines,

(Maldonado-Basc

on et al., 2007), have been the

recognition techniques most widely used. Speciﬁ-

cally, the last one is well known by their excellent

generalization properties as it can be demonstrated

even when the dataset has a high number of classes.

Other approaches are based on the use of Matching

Pursuit (MP) (Hsu and Huang, 2001) and Human Vi-

sion Models (Gao et al., 2006). Although the input

to these classiﬁers is in most cases the gray level of

the pixels in the space domain, other researches con-

form the underlying vector extracting features, such

as histograms (Vicen-Bueno et al., 2005), FFT com-

puted after a complex log-mapping of exterior borders

(Hibi, 1996) and wavelets (Hsu and Huang, 2001).

269

Lafuente-Arroyo S., Maldonado-Bascón S., Gómez-Moreno H. and Gil-Jiménez P..

FALSE ALARM FILTERING IN A VISION TRAFFIC SIGN RECOGNITION SYSTEM - An Approach based on AdaBoost and Heterogeneity of Texture.

DOI: 10.5220/0003156402690276

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 269-276

ISBN: 978-989-8425-40-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

(a) (b) (c) (d)

(e) (f)

Figure 1: Examples of segmentation processes. (a), (c), (e) Original images; (b) red mask; (d) achromatic mask; (f) blue

mask.

2 SYSTEM OVERVIEW

An exhaustive description of the trafﬁc sign detection

and recognition system (TSDRS) on which this re-

search is based can be found in (Maldonado-Basc

et al., 2007) and (Gil-Jim

enez et al., 2008). The sys-

tem consists of the following steps: segmentation, de-

tection and recognition. Although this work is fo-

cused in the improvement of the last stage, it is neces-

sary to make a brief description of the whole system.

2.1 Segmentation

The purpose of this stage is to isolate candidate traf-

ﬁc signs from the background of the scene. Color

information, speciﬁcally HSI space, is considered in

our system to extract candidate objects from the in-

put image by thresholding. Major advantage of HSI

space is that its color components, Hue and Satura-

tion, are closely related to human perception. In ad-

dition, an extra achromatic decomposition similar to

the one used in (Liu et al., 2002) is implemented. In

Figure 1 some segmentation examples are illustrated,

including only the interest masks for each case.

2.2 Detection

The detection block aim is the identiﬁcation of the

shape of each candidate blob. Our algorithm com-

pares the signature of the objects under analysis with

the theorethical signatures in a discrete set of angles,

being deﬁned the signature as the distance from the

mass center to the edge of the blob as a function of the

angle. The Figure 2 shows the signature of a triangle.

r(a)

(a)

p/4 p/2 3p/4 2p

a(rad)

r( )a

0.05

0.1

0.15

0.2

0.25

(b)

Figure 2: Shape signature of a triangular sign. (a) Reference

shape; (b) associated signature.

To make the algorithm invariant to object rotations,

comparisons are performed using the absolute value

of the FFT of the signature.

2.3 Recognition

The purpose of this stage is to identify the informa-

tion related to the pictogram of candidate trafﬁc signs.

This step is implemented by SVM with Gaussian ker-

nel where the input vector is a normalized-size block

in gray-scale for each candidate blob. The strategy

we follow is one-against-all with a number of clas-

siﬁers equal to the number of classes that belong to

each case to analyze, regarding to color and shape.

We must point out that only some pattern vectors of

the training set deﬁne the decision hyperplane. These

pattern vectors are known as support vectors.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

270

3 COMPUTATIONAL LOAD

As well as achieving high accuracy, a recognition sys-

tem should also prevent erroneous identiﬁcation of

non-signs, i.e., limiting the number of false alarms

and even when the purpose is not oriented to real time

applications (in our case we are concerned about road

maintenance tasks), the computation time should be

as low as possible. Computational time required to

process an image in a TSDRS depends on multiple

factors. The most relevant are related to:

1. Image Properties. The properties of the images

to capture are easily conﬁgurable through the ac-

quisition system. Computational load in the seg-

mentation stage is strongly inﬂuenced by image

size, especially when algorithms work in a pix-

elwise fashion. We can reduce the image size

considering a trade-off between speed and detec-

tion probability since small objects in the scene

are difﬁcult to detect and identify. Furthermore,

in the case of a TSDRS that includes tracking it

is crucial to detect the signs since that appear in

the ﬁrst frames of the sequence with small sizes.

Other criterion to consider is wether the system

works with grayscale or color images. Processing

with grayscale images demands lower computa-

tional load but color information is lost. With the

purpose to reduce the image analysis a possible

alternative is to deﬁne the area to explore.

2. Number of Segmentation Algorithms. As we

demonstrate in (G

omez-Moreno et al., 2010)

there’s not an algorithm robust enough to all dif-

ﬁculties from outdoor environments. For this rea-

son, our TSDRS allows us to work with different

algorithms in parallel although their information

is highly redundant and the load complexity in-

creases.

3. Complexity of the Recognition Module. In a

recognition system based on SVMs the number

of support vectors grows as the number of classes

and training samples do.

In order to ﬁnd the main bottlenecks and optimize

the system to improve its performance, we analyze

the computation proﬁle. In Table 1 proﬁles of the

three main stages mentioned are summarized. The

rest of processing time is dedicated to other tasks,

such as image read/write operations. By a simple in-

spection, we can observe that computational load of

the recognition process is approximately 15 and 46

times higher than the corresponding to the detection

stage and segmentation detection stage, respectively.

The reason is a consequence of the high number of

support vectors to manage in the test phase when a

realistic road sign database is considered.

Table 1: Computational load in the three sub-stages of the

TSDRS.

Process CPU cycles

Recognition 49363

Detection 3118

Segmentation 1068

In this way, since the recognition module based on

SVMs is executed for every candidate object, com-

putational cost for each frame increases linearly with

the number of objects at the input of the recognition

stage. Unfortunately, most of theses objects are false

positives. So, in Fig.3 we can observe the output

detection module for an image, to which we apply

two segmentation algorithms. Note that all the de-

tected objects are identiﬁed through their correspond-

ing geometric shape. In the recognition stage all false

alarms are discarded, but our aim in this research is to

reduce the number of objects evaluated in this process

due to its computational load.

In this research our proposal is to decrease the

number of false positives at the input of the recog-

nition module in order to minimize the computational

load. The aim is to implement a false alarm ﬁlter us-

ing the Viola-Jones detector as a previous step to the

recognition module based on SVM and so, reduce the

load of the TSDRS.

4 FALSE ALARM FILTER

In machine learning community it is well known

that more complex classiﬁcation functions yield lower

training errors yet having the risk of poor generaliza-

tion. If the main consideration is test set error, struc-

tural risk minimization provides a formal mechanism

to select a classiﬁer with the right balance of com-

plexity and training error. Another signiﬁcant con-

sideration in classiﬁer design is computational com-

plexity. Since time and error are fundamentally dif-

ferent quantities, no theory can simply select the op-

timal trade-off. Nevertheless, for many classiﬁcation

functions computation time is directly related to the

structural complexity. In this way temporal risk mini-

mization is clearly related to structural risk minimiza-

tion.

This direct analogy breaks down in situations

where the distribution of classes is highly skewed.

For example, in our TSDRS there may be dozens of

false positives among one or two trafﬁc signs in an im-

age. In these cases we can reach high detection rates

and extremely fast classiﬁcations. The key insight is

FALSE ALARM FILTERING IN A VISION TRAFFIC SIGN RECOGNITION SYSTEM - An Approach based on

AdaBoost and Heterogeneity of Texture

271

(a)

(b)

(c)

Figure 3: Candidate objects before the recognition stage.

(a),(b) Candidate objects represented by their correspond-

ing shapes for each segmentation algorithm. (c) Trafﬁc sign

recognition results once detected objects from both algo-

rithms have been integrated.

that while it may be impossible to construct a simple

classiﬁer that can achieve a low training/test error, in

some cases it is possible to construct a simple classi-

ﬁer with a very low false negative rate. For example, it

is relatively simple to construct an extremely fast clas-

siﬁer with a very low false negative rate (i.e. it almost

never misses a sign) and a 50 false positive rate. Such

a detector should be more accurately called a classiﬁ-

cation pre-ﬁlter: when an image region is labeled non-

sign then it can be immediately discarded, but when

a region is labeled sign then further classiﬁcation ef-

fort is required. Such pre-ﬁlter can be implemented

through a cascade of classiﬁers (see Figure 4), where

Figure 4: Schematic diagram of a detection cascade. The

initial stages eliminates a large number of negative exam-

ples with very little processing. After several layers of pro-

cessing the number of false alarms have been reduced radi-

cally.

each classiﬁer follows the AdaBoost algorithm. In the

design of such structure, several parameters have to be

ﬁxed: maximum false positive rate (F

max

) of the cas-

cade, minimum detection rate (D

min

) of the cascade,

maximum false positive rate ( f

max

) of the ith clas-

siﬁer and minimun detection rate (d

min

) of the ith

classiﬁer. Given a trained cascade of classiﬁers, the

detection and false positive rate are:

F =

∏

i=1

(1)

D =

∏

i=1

(2)

where K is the number of classiﬁers, F and D are the

false positive rate and detection rate of the cascade,

respectively, and f

and d

are the false positive rate

and the detection rate, respectively, of the ith classi-

ﬁer.

Even though there are many stages, most are not

evaluated for many noisy samples since they are dis-

carded at the ﬁrst stages. In a cascade, computa-

tion time and detection rate of the ﬁrst stages is crit-

ically important to overall performance. In the cas-

cade structure each stage is implemented according to

the AdaBoost algorithm due to its low computational

complexity in the test phase.

Major extensions of this method have been pro-

posed in two directions: improvement of the algo-

rithm and feature sets.

4.1 Dissociated Dipoles

A more general type of features than the Haar-like

ones, the dissociated dipoles or sticks have been pre-

sented by Balas and Sinha (Balas and Sinha, 2003),

which are composed of a pair of rectangular elements,

named the excitatory dipole and the inhibitory dipole,

respectively (see Fig. 5). As in the case of Haar-like

features, the integral image is used to calculate the

sum of the pixels inside the rectangular regions and

the feature value is equal to the difference between

the values of both dipoles normalized by the number

of stick pixels.

Taking into consideration color information, we

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

272

(a) (b) (c) (d) (e)

Figure 5: Examples of dissociated dipoles. The white re-

gion corresponds to the excitatory dipole whereas the black

region corresponds to the inhibitory region.

(a) (b) (c)

Figure 6: Channels of feature extraction in an example. (a)

Luminance channel, (b) Red Channel, (c) White channel.

propose to apply dipoles in different color compo-

nents. When the dipole is applied to each color chan-

nel, the feature represents the average intensity of a

speciﬁed color component over the region. Specif-

ically, we consider four color channels: red, blue,

white and yellow and an additional luminance chan-

nel. Previously to the integral image, we extract the

color mask in the same way than in segmentation

process using HSI for red and blue channels and an

achromatic decomposition for white channel. In Fig.6

we can observe the masks corresponding to the lumi-

nance and the red and achromatic channels for one ex-

ample. The blue and yellow masks are not illustrated

since all pixels have a zero value.

The contribution of a dissociated dipole for each

channel is computed as the difference mentioned

above between the excitatory and inhibitory regions:

Value =



channel



exc

−



channel



inh

(3)

In the case of the luminance channel, I

channel

rep-

resents the sum of gray levels in region delimited by

the stick, whereas in each color channel represents the

number of pixels segmented by this color. N

channel

the number of pixels of the region which are covered

by the stick.

With the purpose to make qualitative comparisons

between the sticks, we can simplify the expression

considering only the sign of the difference. This ap-

proach presents luminance normalization as advan-

tage.

4.2 Discrete AdaBoost

AdaBoost algorithm ﬁnds precise hypotheses by com-

bining several weak classiﬁcation functions which

have moderate precision. AdaBoost is an iterative al-

gorithm that ﬁnds, from a feature set, some weak but

discriminative classiﬁcation functions and combines

them in a strong classiﬁcation function:

H =







∑

t=1

≥

∑

t=1

= S,

0, otherwise

(4)

where H and h

are the strong and weak classiﬁcation

functions, respectively, and α

is a weight coefﬁcient

for each h

. Different variants of boosting are devel-

oped: discrete AdaBoost, real AdaBoost, gentle Ad-

aBoost, and so forth. However, we use the ﬁrst one.

Each weak classiﬁer is deﬁned for a feature j as a

binary response:



1, ifp

< p

0, otherwise

(5)

where f

is the value of the feature j, which is

given by the contribution of a dipole as it is men-

tioned above, θ

is the threshold and p

is the parity.

For each feature j, AdaBoost determines an optimal

threshold θ

for which the classiﬁcation error on train-

ing database (with positive and negative examples) is

minimized. The weight coefﬁcient for each h

is com-

puted as:



1 −e



(6)

where e

is the sum of the weigths associated to the

samples classiﬁed wrongly.

4.3 Assymetric Recognition

Asymmetry is an inherent problem in recognition sys-

tems where the number of positive targets is too low

with respect to negative patterns. That is, to achieve a

high detection rate, the cost of missing a target should

be higher than that of a false positive. Cost-sensitive

learning is a suitable way for solving such problems.

However, most cost-sensitive extensions of AdaBoost

are realized by heuristically modifying the weights

and conﬁdence parameters of the discrete AdaBoost.

Thus, there should be different treatment for false

negatives (FN) and false positives (FP), that is, FN

samples are penalized more than FP samples. Since

AdaBoost aims at minimizing the bound of classiﬁca-

tion error which treats FP and FN equally, the sym-

metric AdaBoost algorithm is not optimal for object

detection tasks. To deal with the class imbalance

problem in classiﬁcation, various asymmetric exten-

sions of AdaBoost have been proposed in the litera-

ture. Most of them directly modify the weights and

conﬁdence parameters of discrete AdaBoost with-

out clarifying the relations to the loss minimization

FALSE ALARM FILTERING IN A VISION TRAFFIC SIGN RECOGNITION SYSTEM - An Approach based on

AdaBoost and Heterogeneity of Texture

273

Figure 7: Examples of the benchmark data set. First row

includes positive samples whereas second and third row in-

clude noisy ones.

of AdaBoost. AdaCost (Fan et al., 1999) proposed

by Fan adopts an approach to make AdaBoost cost-

sensitive. They incorporated a cost adjustment func-

tion β

(i) into the weight updating rule and the com-

putation of δ

. The weight updating formula was

modiﬁed into

t+1

(i) =

(i)exp(−α

)β

(i))

(7)

where α

is computed as:



1 +r

1 −r



(8)

and r

∑

(i)y

)β

(i) (9)

β=+1 if the output of the classiﬁer is right and β=-1

otherwise. Finally,

(i) = −0.5C

+ 0.5 (10)

and

−1

(i) = 0.5C

+ 0.5 (11)

The parameter C

is the cost factor assigned to the

i-th sample and is restricted to the interval [0 1].

4.4 Results

The Recognition and Multi-sensorial analysis group

(GRAM) at the Universidad de Alcal

a has collected

a complete database of Spanish trafﬁc signs. All the

samples have been extracted from images acquired by

different video-cameras under variable lighting condi-

tions. The stored patterns are 31×31 pixels gray level

with homogeneous background for no-interest pixels.

So, the number of signiﬁcative components is 961 for

rectangular signs while in circular signs is reduced to

709 and for for triangular signs to 511. Some exam-

ples are shown in Figure 7.

Without loss of generality, in this research we are

concerned about the speed limit signs. The bench-

mark data set is subdivided into three groups: train-

ing, validation and test set. From a total set of 872

positive samples and 3475 negative samples, 50% of

them are chosen randomly as training set, 25% for

the validation set and the remaining 25% forming the

test set. Negative samples have been previously taken

randomly from arbitrary images. The validation set is

used to tune the strong classiﬁer decision thresholds

in order to reach the minimum acceptable correct de-

tection rate and the maximum acceptable false-alarm

rate during cascade training.

In the ﬁnal detector the selected values for false

positive rate and detection rate of each strong classi-

ﬁer are, respectively, ﬁxed to f

max

= 0.9 · f

i−1

max

and d

min

= 0.994 · f

i−1

min

. Moreover, the false pos-

itive rate and detection rate of the cascade are ﬁxed

to F=0.05 and D=0.9. The structure of the ﬁnal de-

tector is a 11 layer cascade of classiﬁers with a total

of 28 features. In Table 2 is summarized the distri-

bution of features among the different stages. The

ﬁlter achieves a detection rate of 98.58% (3 positive

samples of 211 were considered false negatives) and

a false positive rate of 5.30% (822 negative samples

of 868 were discarded).

In the case of AdaCost boosting, using the same

parameters for the cascade structure, the detector we

obtained has 6 layers with a total de 28 features. In

Table 3 is summarized the distribution of features

among the different stages. The ﬁlter achieves a de-

tection rate of 98.10% and a false positive rate of

4.15%.

Anyway, from the Figure 7, we can observe that

pictograms of signs are well deﬁned. We can take

advantage of this fact binarizing the samples using

Otsu’s method, which chooses a global threshold to

minimize the intraclass variance of the thresholded

black and white pixels. In Figure 8 are illustrated the

samples of Fig 7 after thresholding. For each bina-

rized image we can quantify the heterogeneity of tex-

ture computing for every pixel of coordinates (x,y) the

number the pixels in 8-neighboring that present differ-

ent level respect to it. The value is normalizazed with

respect to the total number of pixels that constitute the

object.

Normalized mean value of heterogeneity is shown

by the solid line in Figure 9, whereas drifts (maximun

and minimun) are shown by the dashed lines for the

different classes. The last label (Index = 10) corre-

sponds to the noisy class. Note that the distribution

shows a higher disparity of values for noisy samples

(class 10) and it may be discriminant enough for many

cases. Thus, only those candidate objects that present

a heterogeneity lower than θ

= 0.75 and higher than

> 1.80 will be analyzed by the cascade ﬁlter. In

fact, 411 negative samples of the 868 (47.35%) that

conform the negative test set were discarded using

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

274

Table 2: Structure, detection rates and false alarm rates of the ﬁlter, which includes 11 boosting layers.

L1 L2 L3 L4 L5 L6 L7 L8 L9 L10 L11

No.Features 3 2 2 2 2 4 2 2 3 3 3

Detection Rate 99.0% 99.0% 99.0% 98.5% 98.5% 98.5% 98.5% 98.5% 98.5% 98.5% 98.5%

FA Rate 46.7% 31.2% 26.2% 21.3% 19.1% 16.9% 14.0% 11.0% 8.9% 7.2% 5.3%

Table 3: Structure, detection rates and false alarm rates of the ﬁlter based on AdaCost boosting, which includes 6 stages.

L1 L2 L3 L4 L5 L6

No.Features 7 2 5 10 3 1

Detection Rate 100% 100% 99.5% 99.5% 98.1% 98.1%

FA Rate 56.3% 25.9% 11.8% 8.7% 6.5% 4.1%

Figure 8: Binarized samples of the benchmark data set us-

ing Otsu’s method.

Figure 9: Illustration of the response of heterogeneity of

texture for different classes. The solid graph represents the

mean value whereas the dashed ones depict the minimum

and maximum values. The distribution shows a higher dis-

parity of values for noisy samples (class 10) than for the

rest.

this prerequisite. If we integrate this condition in the

cascade ﬁlter detection based on AdaCost, the ﬁlter

maintains a detection rate of 98.10% but the false pos-

itive rate is decreased until 2.65%. Note than in this

case, almost half of noisy examples are discarded di-

rectly by the condition of heterogeneity of texture be-

fore ﬁltering. Furthermore, the false alarm ratio is

improved due to the heterogeneity texture.

Although cascades of boosted ensembles exhibit

real-time performance, training time ranges from days

to weeks. The factors that affect the training time are

the amount of training samples and feature-set size

that depends on the actual image dimensions of train-

ing samples. The traditional training approach has a

run-time of O(N · T · log(N)), where N represents the

number of samples and T is the number of features.

In our case, we work with a Sun Java WorkStation

and the training requires approximately 12 hours of

computation for 1600 features.

5 CONCLUSIONS

In this paper we present a false alarm ﬁlter based on a

cascade of boosted classiﬁers. The ﬁlter is integrated

in a vision system to detect and recognize trafﬁc signs.

In this application the number of positive targets is

too low with respect to negative patterns and the com-

putational load of recognition task depends directly

on the number of samples to identify. In order to

decrease the number of candidate signs to be recog-

nized, the ﬁlter has the capacity of discard many false

positives with very low operations using the structure

of Viola-Jones. One of the main goals of this ﬁlter

is to include a ﬁrst discriminant classiﬁer based on a

measurement of heterogeneity texture. By combining

both techniques, the system achieves a very low false

alarm rate with low computational load.

Since in this work, we are concerned about speed

limit signs, our future work will be devoted to imple-

ment this approach with the whole set of trafﬁc sign

set. Furthermore, the enormous time that Viola-Jones

algorithm demands in training process, makes it an

open item to research.

REFERENCES

Bahlmann, C., Zhu, Y., Ramesh, V., Pellkofer, M., and

Koehler, T. (2005). A system for trafﬁc sign detection,

tracking, and recognition using color, shape, and mo-

tion information. In Intelligent Vehicles Symposium,

Proceedings, IEEE, pages 255–260.

Balas, B. and Sinha, P. (2003). Sticks: Image-representation

via non-local comparisons. Journal of Vision, 3(9).

FALSE ALARM FILTERING IN A VISION TRAFFIC SIGN RECOGNITION SYSTEM - An Approach based on

AdaBoost and Heterogeneity of Texture

275

de la Escalera, A., Armingol, J., and Mata, M. (2003). Traf-

ﬁc sign recognition and analysis for intelligent vehi-

cles. Image and Vision Computing, 21:247–258.

de la Escalera, A., Moreno, L., Salichs, M., and Armingol,

J. (1997). Road trafﬁc sign detection and classiﬁca-

tion. Industrial Electronics, IEEE Transactions on,

44(6):848–859.

Fan, W., Stolfo, S., Zhang, J., and Chan, P. (1999). Adacost:

misclassiﬁcation cost-sensitive boosting. In Proceed-

ings of International Conference on Machine Learn-

ing, pages 97–105.

Fang, C., Chen, S., and C., F. (2003). Road sign detection

and tracking”. IEEE Trans. on Vehicular Technology,

52(5):1329–1341.

Gao, X., Podladchikova, L., Shaposhnikov, D., Hong, K.,

and Shevtsova, N. (2006). Recognition of trafﬁc signs

based on their colour and shape features extracted us-

ing human visual models. Journal of Visual Commu-

nication and Image Representation, 17(4):675–685.

Gil-Jim

enez, P., Maldonado-Basc

on, S., G

omez-Moreno,

H., Lafuente-Arroyo, S., and L

opez-Ferreras, F.

(2008). Trafﬁc sign shape classiﬁcaction and localiza-

tion based on the normalized FFT of the signature of

blobs and 2D homographies. Elsevier Science. Signal

Processing, 88(2943):2955.

omez-Moreno, H., Maldonado-Basc

on, S., Gil-Jim

enez,

P., and Lafuente-Arroyo, S. (2010). Goal-evaluation

of segmentation algorithms for trafﬁc sign recogni-

tion. IEEE Trans. on Intelligent Transportation Sys-

tems. Accepted for publication.

Hibi, T. (1996). Vision based extraction and recognition

of road sign region from natural color image, by us-

ing hsl and coordinates transformation. In Proc. of

the 29th. Int. Symposium Automotive Technology and

Automation, Florence, Italy.

Hsu, S. H. and Huang, C. L. (2001). Road sign detection

and recognition using matching pursuit method. Im-

age and Vision Computing, 19:119–129.

Liu, H., Liu, D., and Xin, J. (2002). Real-time recognition

of road trafﬁc sign in motion image based on genetic

algorithm. Proc. of the 1st. Int. Conference on Ma-

chine Learning and Cybernetics, pages 83–86.

Maldonado-Basc

on, S., Lafuente-Arroyo, S., Gil-Jim

enez,

P., G

omez-Moreno, H., and L

opez-Ferreras, F. (June

2007). Road-sign detection and recognition based on

support vector machines. Intelligent Transportation

Systems, IEEE Transactions on, 8(2):264–278.

Vicen-Bueno, R., Gil-Pita, R., Rosa-Zurera, M., M. Utrilla-

Manso, M., and L

opez-Ferreras, F. (2005). Multilayer

perceptrons applied to trafﬁc sign recognition tasks.

In Proc. of the 8th International Work-Conference on

Artiﬁcial Neural Networks,IWANN, pages 865–872.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

276